American Journal of Epidemiology

doi:10.1093/aje/153.4.345
153:345-352, 2001. Am. J. Epidemiol.
Marshall
Theodore G. Liou, Frederick R. Adler, Stacey C. FitzSimmons, Barbara C. Cahill, Jonathan R. Hibbs and Bruce C.
Predictive 5-Year Survivorship Model of Cystic Fibrosis
Supplement/Special Issue
http://aje.oxfordjournals.org/cgi/content/full/153/4/345/DC1
"Replacements for Johns Hopkins JEPI Web page" This article is part of the following issue:
http://aje.oxfordjournals.org/cgi/content/full/153/4/345
The full text of this article, along with updated information and services is available online at
References
http://aje.oxfordjournals.org/cgi/content/full/153/4/345#BIBL
This article cites 38 references, 16 of which can be accessed free at
Cited by
http://aje.oxfordjournals.org/cgi/content/full/153/4/345#otherarticles
This article has been cited by 53 articles at 14 May 2008 . View these citations at
Supplementary material
http://aje.oxfordjournals.org/cgi/content/full/153/4/345/DC1
Data supplements for this article are available at
Reprints http://aje.oxfordjournals.org/cgi/reprintform Reprints of this article can be ordered at
Email and RSS alerting
http://aje.oxfordjournals.org
Sign up for email alerts, and subscribe to this journals RSS feeds at
image downloads
PowerPoint Images from this journal can be downloaded with one click as a PowerPoint slide.
Journal information
subscribe can be found at http://aje.oxfordjournals.org
Additional information about American Journal of Epidemiology, including how to
Published on behalf of
http://www.oxfordjournals.org
Oxford University Press

at Australian National University on 14 May 2008 http://aje.oxfordjournals.org Downloaded from
345
American Journal of Epidemiology
Copyright 2001 by The Johns Hopkins University School of Hygiene and Public Health
All rights reserved
Vol. 153, No. 4
Printed in U.S.A.
Five-Year Survivorship Model of Cystic Fibrosis Liou et al.
Predictive 5-Year Survivorship Model of Cystic Fibrosis
Theodore G. Liou,
1,2
Frederick R. Adler,
3,4
Stacey C. FitzSimmons,
5,9
Barbara C. Cahill,
1,2,6
Jonathan R. Hibbs,
7
and Bruce C. Marshall
1,2,8
The objective of this study was to create a 5-year survivorship model to identify key clinical features of cystic
fibrosis. Such a model could help researchers and clinicians to evaluate therapies, improve the design of
prospective studies, monitor practice patterns, counsel individual patients, and determine the best candidates
for lung transplantation. The authors used information from the Cystic Fibrosis Foundation Patient Registry
(CFFPR), which has collected longitudinal data on approximately 90% of cystic fibrosis patients diagnosed in
the United States since 1986. They developed multivariate logistic regression models by using data on 5,820
patients randomly selected from 11,630 in the CFFPR in 1993. Models were tested for goodness of fit and were
validated for the remaining 5,810 patients for 1993. The validated 5-year survivorship model included age, forced
expiratory volume in 1 second as a percentage of predicted normal, gender, weight-for-age z score, pancreatic
sufficiency, diabetes mellitus, Staphylococcus aureus infection, Burkerholderia cepacia infection, and annual
number of acute pulmonary exacerbations. The model provides insights into the complex nature of cystic fibrosis
and supplies a rigorous tool for clinical practice and research. Am J Epidemiol 2001;153:34552.
cystic fibrosis; logistic models; models, theoretical; multivariate analysis; proportional hazards models; survival
analysis
Received for publication June 30, 2000, and accepted for publi-
cation November 7, 2000.
Abbreviations: CFFPR, Cystic Fibrosis Foundation Patient
Registry; FEV
1
, forced expiratory volume in 1 second; FEV
1
%, forced
expiratory volume in 1 second as a percentage of predicted normal.
1
Division of Respiratory, Critical Care and Occupational
Pulmonary Medicine, Department of Internal Medicine, Health
Sciences Center, University of Utah, Salt Lake City, UT.
2
The Intermountain Cystic Fibrosis Center, University of Utah,
Salt Lake City, UT.
3
Department of Mathematics, University of Utah, Salt Lake City,
UT.
4
Department of Biology, University of Utah, Salt Lake City, UT.
5
Cystic Fibrosis Foundation, Bethesda, MD.
6
Lung Transplantation Program, Health Sciences Center,
University of Utah, Salt Lake City, Utah.
7
The New York State Department of Health, Albany, NY.
8
The Salt Lake Veterans Administration Medical Center, Salt
Lake City, UT.
9
Present address: FitzSimmons and Associates, Bethesda, MD.
Reprint requests to Dr. Theodore G. Liou, Division of Respiratory,
Critical Care and Occupational Pulmonary Medicine, University of
Utah, 50 North Medical Drive, Salt Lake City, UT 84132 (e-mail:
ted.liou@m.cc.utah.edu).
Cystic fibrosis is an autosomal recessive, multisystem
disease leading to significant morbidity and early death.
Since the disease was described in 1938 (1, 2), treatments
for its pancreatic and pulmonary manifestations have
improved median survival in the United States from less
than 6 months to about 32 years in 1998 (3). Severe pul-
monary disease is the primary cause of cystic-fibrosis-
related mortality, constituting 76.4 percent of such deaths
in 1998 (3). Chronic inflammation and infection of the air-
ways characterize cystic-fibrosis-related pulmonary dis-
ease. Staphylococcus aureus and Haemophilus influenzae
infections, commonly found early in the course of the dis-
ease, are often supplanted by Pseudomonas aeruginosa
infection as the disease progresses (3). Infection with
Burkerholderia cepacia is associated with accelerated
pulmonary disease (4, 5). Malnutrition, in part due to
pancreatic insufficiency, was the major feature of the dis-
ease according to early reports, and it continues to be a
substantial problem (13). With improved survival, addi-
tional manifestations such as diabetes mellitus have been
recognized (6, 7).
Many studies have considered the survival effect of a
variety of clinical and physiologic features of cystic fibro-
sis such as forced expiratory volume in 1 second as a per-
centage of predicted normal (FEV
1
%), gender, age, preg-
nancy, or particular therapies (426). We developed a single
survivorship model that integrated many characteristics of
cystic fibrosis and quantified the relative contribution of
each.
The current most commonly used survival model of cys-
tic fibrosis was developed in 1992 and is based on FEV
1
%
alone or on age, gender, and FEV
1
% (15). Clinicians often
use this model to select patients to refer for lung transplan-
tation. The model is relatively simple to use for estimating
cystic fibrosis survival, but it has not been validated and
does not incorporate clinical features of cystic fibrosis now
recognized as important predictors of mortality (47, 9, 12,
14, 15, 18, 2226). We examined the validity of the previ-
ous simple models and compared their accuracy with that of
our new model.
346 Liou et al.
Am J Epidemiol Vol. 153, No. 4, 2001
MATERIALS AND METHODS
Cystic Fibrosis Foundation Patient Registry data
We used information from the Cystic Fibrosis Foundation
Patient Registry (CFFPR; 19861997), which contains lon-
gitudinal data on 27,849 patients at 115 cystic fibrosis care
centers, representing approximately 90 percent of all cystic
fibrosis patients in the United States (3). Reports for each
patient, containing a wide range of clinical data, are submit-
ted annually to the CFFPR. CFFPR methods are described
elsewhere (27). In accordance with the Cystic Fibrosis
Foundation data access policy, we applied for and received
access to a 19861996 longitudinal database, which was
later updated with the complete 1997 database and 1998
mortality data.
Patient selection
Cystic fibrosis patients who were alive on January 1,
1993, and for whom follow-up data were available through
December 31, 1997, were included in our study. Data on
patients with odd identification numbers were used to
develop survival models; those from patients with even
identification numbers were used for the validation set.
Patients were excluded from the study if they had undergone
solid organ transplantation of any type or lacked pulmonary
function testing information, a key clinical parameter of cys-
tic fibrosis.
Data handling and statistical modeling tools
We used S-PLUS software (version 3.4, release 1 for Sun
SPARC, SunOS 5.3: 1996 (Mathsoft, Inc., Cambridge,
Massachusetts)) to perform statistical analyses (28). S-PLUS
is hosted by the Sun UltraSPARC 2170/2200 cluster (Sun
Microsystems, Inc., Palo Alto, California) of the University
of Utah Department of Mathematics.
Prediction equations for FEV
1
% and forced vital capacity
Raw spirometry values, forced expiratory volume in 1
second (FEV
1
), and forced vital capacity were normalized to
FEV
1
% and percentage of predicted forced vital capacity.
We used regression formulae from the Third National
Health and Nutrition Examination Survey (29).
Calculation of weight-for-age z scores
Weight-for-age z scores were calculated by using
National Center for Health Statistics (Hyattsville,
Maryland) weight-for-age growth tables for males and
females. Gender and age were used to determine the
appropriate median weight for age for each patient in
the 1993 CFFPR. Because the distribution of weight for
age was not normal, the z score was calculated by using
the approximation method of Lai et al. and Dibley et al.
(13, 30).
Rate of decline in FEV
1
%
To estimate the slopes of the rates of decline in FEV
1
%,
linear and mixed-effects linear models were generated for
each patient who was alive on January 1, 1993, and for
whom there was more than one measurement (23). We used
all 19861992 pulmonary function test data (two to five
measurements).
Statistical models for survival analysis
To predict 5-year survivorship for cystic fibrosis patients,
we used a binary variable, alive or deceased within a 5-year
time period, as the outcome variable. All potential explana-
tory covariates were incorporated into a logistic regression
model. Forward stepwise procedures and log-likelihood
ratio tests were used to select variables for the model.
Parallel analysis with Cox proportional hazards regression
was performed.
Selection of variables for survival analysis
Data on 317 variables were collected by the CFFPR in
1993. We excluded variables that related directly to ongoing
research protocols, for which data were sparse, or that
recorded clinically uncommon or rare characteristics (table
1). However, our analysis included infection with B. cepa-
cia (2.8 percent of the population affected) because of prior
reports of its effects on mortality (4, 5, 24, 31) and infection
with Stenotrophomonas maltophilia (1.7 percent of the pop-
ulation affected) because of its increasing prevalence (3).
The pertinent variables, including weight-for-age z scores
and FEV
1
% slopes, were tested by using univariate logistic
regression for an effect on 5-year survivorship. Potentially
interacting variables were identified by using stratified
regression analysis (31).
Modeling of 5-year survivorship
Variables that demonstrated an association with 5-year sur-
vivorship at an absolute t-test value of more than two were
considered for multivariate logistic regression analysis.
Twenty-one variables were included in the final multivariate
analysis. The model was subjected to stepwise analysis of
variance to derive an intermediate model (32). By using these
variables, along with patient age, we constructed models that
included all possible interactions and excluded statistically
nonsignificant terms by using stepwise analysis of variance
prior to validation. All possible interaction terms were con-
sidered, including those identified by stratified analysis, and
terms that lacked statistical significance were excluded (31).
For comparison, we developed simpler models by using only
the covariates identified by Kerem et al. (15).
Assessing models for goodness of fit and validation
Models developed by using information from odd-
numbered patients were validated by using 1993 patients
who had even identification numbers. We used the Hosmer-
Five-Year Survivorship Model of Cystic Fibrosis 347
Am J Epidemiol Vol. 153, No. 4, 2001
Lemeshow test to eliminate models that fit poorly; the
remaining models were compared with standard
2
likeli-
hood ratio statistics (31).
RESULTS
The CFFPR contains 1993 data on 19,156 patients. Of
these, we excluded 730 patients because they were recipi-
ents of solid organ transplantation and 5,686 because they
lacked FEV
1
% measurements. Of these latter patients, 4,190
were excluded because they were younger than age 5.5
years, a group for whom FEV
1
% cannot be measured using
standard techniques, or were members of ethnic or racial
groups for whom FEV
1
% standards do not exist (54
patients). Microbiology, pancreatic sufficiency, diabetes, or
acute exacerbation information was missing for 1,056
patients. Of the 12,686 patients remaining after exclusions,
we included information on 11,630 (92 percent) of them in
developing and validating the model.
We developed the survivorship model by using data on
5,820 patients with odd identification numbers, and we val-
idated the model externally by using data on 5,810 patients
with even identification numbers. Baseline characteristics of
the development and validation patient populations were
similar (table 2). There were 1,419 deaths spread evenly
between the development and validation groups.
Univariate logistic regression analyses identified 21 vari-
ables with a significant predictive value for 5-year survivor-
ship. Clinically relevant variables that did not add predictive
information to the model were eliminated (table 1). In our
investigation of interactions among the remaining variables
using stratified analysis, we found five potential interaction
terms for multivariate analysis.
The best multiple logistic regression model included nine
variables with one interaction (table 3). Higher FEV
1
%,
higher weight-for-age z score, pancreatic sufficiency, and S.
aureus infection predicted increased survivorship.
Increasing age, female gender, diabetes mellitus, B. cepacia
infection, and a higher number of acute pulmonary exacer-
bations predicted decreased survivorship. After stepwise
analysis, only one of the five potential interaction terms (B.
cepacia infection number of acute pulmonary exacerba-
TABLE 1. Covariates excluded from the survivorship model used to identify key clinical features of
cystic fibrosis, Cystic Fibrosis Foundation Patient Registry, United States, 1993
Not statistically significant* Insufficient data
Height (raw value and percentile)
F508
Pseudomonas aeruginosa
Stenotrophomonas maltophilia
Haemophilus influenzae
Pregnancy
Sweat chloride test value at diagnosis
Education
FEV
1
% slope by linear regression
FEV
1
% slope by mixed-effects modeling
Other Pseudomonas
Mucoid Pseudomonas
Alcaligenes xylosoxidans
Aspergillus fumigatus
Methillicin-resistant Staphylococcus aureus
Birth complications
Employment
Insurance status
Marital status
* p > 0.05.
Information on some of these covariates was available from less than 5 percent of the population. If
information was available, covariates were excluded if less than 5 percent of the cystic fibrosis population was
affected; most of these excluded variables affected fewer than 10 patients.
A specific test of whether the presence of no, one, or two F508 alleles had a survival effect.
FEV
1
%, forced expiratory volume in 1 second as a percentage of predicted normal.
TABLE 2. Characteristics of model development and
validation patients, Cystic Fibrosis Foundation Patient
Registry, United States, 1993
Characteristic
Patients
No. of patients
Mean age (years) (SD*)
Median age (years) (range)
Mean FEV
1
%* (SD)
Female gender (%)
Mean weight-for-age z score (SD)
Pancreatic sufficiency (%)
With diabetes mellitus (%)
With Staphylococcus aureus (%)
With Burkerholderia cepacia (%)
Mean no. of acute exacerbations
(range)
5,810
17.64 (9.36)
15.32
(5.571.05)
67.4 (29.3)
2,692 (46)
0.84 (1.05)
308 (5.3)
367 (6.3)
1,802 (31.0)
231 (4.0)
1.2
(019)
* SD, standard deviation; FEV
1
%, forced expiratory volume in 1
second as a percentage of predicted normal.
5,820
17.58 (9.25)
15.52
(5.562.44)
67.7 (29.2)
2,729 (47)
0.85 (1.07)
311 (5.3)
357 (6.1)
1,780 (30.6)
186 (3.2)
1.1
(018)
Development Validation
348 Liou et al.
Am J Epidemiol Vol. 153, No. 4, 2001
tions) was statistically significant. The single interaction
term showed that each acute exacerbation had a markedly
reduced, although persistent, negative effect when B. cepa-
cia infection was present.
We normalized the effect of each variable in terms of the
equivalent loss or gain in FEV
1
% to show the relative effect
of each variable in a clinically meaningful way (table 3). For
example, B. cepacia infection was a major negative predic-
tor of survival, as expected on the basis of previous reports
(4, 5, 24). The survival effect was the same as a 48 percent
drop in FEV
1
% in a patient without B. cepacia infection.
Four acute exacerbations in a single year had the same neg-
ative effect on survival as infection with B. cepacia.
Parallel analysis in which we used Cox proportional haz-
ards regression generated a model with identical covariates
and similar coefficients (table 4). Because the logistic
regression model is potentially simpler to use and interpret
in a clinical setting, we focused our analysis on the results
of logistic regression modeling (33, 34).
Simpler models using age, gender, and FEV
1
% or FEV
1
%
alone, the covariates identified by Kerem et al. (15), pro-
duced much lower log-likelihood scores than the new model
(table 5) did. The simpler models failed the Hosmer-
Lemeshow test for goodness of fit on the validation set of
data (31) and failed to identify the patients with the lowest
5-year conditional probability of survival (figure 1).
DISCUSSION
We developed and validated a 5-year survivorship model
of cystic fibrosis that identified eight characteristics of the
disease, in addition to FEV
1
%, that together accurately pre-
dict survival. The model provides insight into the relative
effect of each characteristic and underscores the importance
of considering multiple clinical factors when assessing the
likelihood of 5-year survival. It is generalizable to cystic
fibrosis patients who have undergone pulmonary function
testing. In addition, it may provide a method of estimating
TABLE 3. Validated 5-year logistic regression survivorship model for cystic fibrosis,* Cystic Fibrosis
Foundation Patient Registry, United States, 1993
Covariate
(x
010
)
Coefficient
(Intercept)
Age (per year)
Gender (male = 0, female = 1)
FEV
1
% (per %)
Weight-for-age z score
Pancreatic sufficiency (0 or 1)
Diabetes mellitus (0 or 1)
Staphylococcus aureus (0 or 1)
Burkerholderia cepacia (0 or 1)
No. of acute exacerbations (05)
No. of acute exacerbations
B. cepacia
* Hosmer-Lemeshow p value = 0.54; no significant difference between predicted and actual survivorship of the
validation group of patients (31).
The conditional probability of 5-year survival by logistic regression analysis is = exp(x)/(1 + exp(x), where
the logit x is:
x =
0
+
1
Age +
2
Gender +
3
FEV
1
% +
4
z score +

5
PancreaticSufficiency +
6
diabetes +

7
S. aureus +
8
B. cepacia +
9
Exacerbations +

10
(Exacerbations B. cepacia)
For a covariate x
i
and its coefficient
i
, the term exp(
i
x
i
) gives the incremental odds ratio for that covariate.
Coefficients for each covariate are unitless unless specified.
FEV
1
%, forced expiratory volume in 1 second as a percentage of predicted normal; SE, standard error.
FEV
1
% equivalence was calculated as
i i
/
FEV1%
. Except for age, values are rounded to the nearest integer.
For age, weight-for-age z score, and acute exacerbations, the equivalences are per year, per z score point, and
per exacerbation, respectively.
FEV
1
% was calculated by using raw FEV
1
results reported to the Cystic Fibrosis Foundation Patient Registry
in 1993 for patients who did not undergo transplantation. Pancreatic sufficiency was defined as not using
pancreatic enzyme supplementation. Diabetes was defined as the need for insulin during or before 1993. Values
for S. aureus and B. cepacia infections are from reported microbiologic data. Number of acute exacerbations is
the number of episodes of acute pulmonary exacerbations of cystic fibrosis requiring treatment in 1993, up to a
maximum of five.
1.93
0.028
0.23
0.038
0.40
0.45
0.49
0.21
1.82
0.46
0.40
Odds
ratio
SE
0.27
0.0060
0.10
0.0028
0.053
0.31
0.15
0.12
0.30
0.031
0.12
6.88
0.97
0.79
1.04
1.50
1.58
0.61
1.24
0.16
0.63
1.49
50
0.7
6
1
10
12
13
6
48
12
10

010
FEV
1
%
equivalence
Five-Year Survivorship Model of Cystic Fibrosis 349
Am J Epidemiol Vol. 153, No. 4, 2001
the impact of therapies that have not been studied in a ran-
domized and controlled fashion. Five-year survivorship of
cystic fibrosis patients may be changing because of changes
in practice patterns, and our model provides a way to gauge
this survival effect. For individual patients, the model
enables prognosis to be estimated.
Previous modeling of cystic fibrosis survivorship concen-
trated on FEV
1
% as the main predictor of 2-year survival
(15). However, this simple model and models based on
FEV
1
%, gender, and age do not pass validation tests and fail
to identify those patients at the highest risk of death, because
they ignore other critical covariates that influence mortality
(47, 9, 12, 14, 15, 18, 2226).
The new, validated model that we developed identified
the additional covariates needed for a predictive survivor-
ship model of cystic fibrosis. Some of these covariates were
not surprising (14, 15, 25). Nutritional status, as indicated
by weight-for-age z score, substantially affects long-term
outcome. Our model confirms that female gender is associ-
ated with a worse long-term outcome (14). The underlying
reason remains unexplained, but we found that pregnancy
was not the cause (9, 11, 35). S. aureus infection was asso-
ciated with improved 5-year survival. We also found the sur-
vival effect independent of age, but the specific underlying
cause is unknown. Each acute pulmonary exacerbation
within the year had an unexpectedly large, negative impact
on 5-year survival equal to subtracting 12 percent from the
measured FEV
1
% value.
Infection with B. cepacia had the largest effect of any
model variable for predicting 5-year survivorship. Cystic
fibrosis patients infected with B. cepacia tend to have the
poorest predicted 5-year survival and often are deemed inel-
igible for lung transplantation because of reports of poorer
post-transplantation outcomes (22, 36).
As median survival continues to increase, diabetes melli-
tus becomes a greater clinical concern. This disease is found
in a minority of children with cystic fibrosis, but its preva-
lence rises to approximately 50 percent by age 30 years and
to 70 percent by age 40 years (10). For cystic fibrosis
patients, diabetes is associated with reduced pulmonary
function (26), increased effort to breathe, and increased
energy expenditure (37). Untreated diabetes may predispose
patients to additional infections (26). The survival effect of
diabetes mellitus in cystic fibrosis patients has been contro-
versial (38), but our model demonstrated a large negative
effect of diabetes on 5-year survival, independent of other
covariates.
Predictive models of disease, similar to ours, are used in
multiple ways as research and clinical tools. One example that
continues to be enormously useful more than 15 years after it
was developed is the multivariate logistic regression model of
the acute physiology and chronic health evaluation
(APACHE) II score (39). These scores are used to correct for
differences in expected mortality between control and exper-
imental groups (40), to weight cost analyses of critical illness
(41), to assess the impact of changes in the health delivery
system (42), and to estimate individual patient prognosis (43).
To facilitate practical and efficient use of the new model
in clinical settings, we developed two worksheets that
TABLE 4. Cox proportional hazards model of cystic fibrosis
developed from Cystic Fibrosis Foundation Patient Registry
data, United States, 1993
Covariate*
(x
010
)
Coefficient
Age (per year)
Gender (male = 0, female = 1)
FEV
1
% (per %)
Weight-for-age z score
Pancreatic sufficiency (0 or 1)
Diabetes mellitus (0 or 1)
Staphylococcus aureus (0 or 1)
Burkerholderia cepacia (0 or 1)
No. of acute exacerbations (05)
No. of acute exacerbations
B. cepacia
0.0049
0.074
0.0025
0.041
0.23
0.098
0.09
0.19
0.024
0.06
* Covariates developed for this model are identical to those for the
logistic regression model (table 3). Values of the covariates are similar but oppo-
site in sign because the Cox model predicts mortality, whereas the logistic
regression model (table 3) predicts survival. Data used to develop this model
are identical to those for the development data set for the logistic regression
model. Unless specified, coefficients for each covariate are unitless.
SE, standard error; FEV
1
%, forced expiratory volume in 1 second as a
percentage of predicted normal.
0.011
0.15
0.042
0.28
0.14
0.44
0.25
1.41
0.35
0.28
Odds
ratio
SE
1.011
1.16
0.96
0.75
0.87
1.55
0.78
4.12
1.42
0.75

010
TABLE 5. Statistical tests of 5-year logistic regression
survivorship models of cystic fibrosis developed from Cystic
Fibrosis Foundation Patient Registry data, United States, 1993
Test of validation
Best model
Best model with no interaction
terms
Model with interaction terms
Age, gender, FEV
1
%
FEV
1
% only#
* The Hosmer-Lemeshow p value calculation (31) measures
how well the predictions of a model fit actual outcomes. A p value of
>0.05 shows that model predictions cannot be distinguished from
actual outcomes and demonstrates successful external validation of
a logistic regression model. These values are calculated
independently of the
2
likelihood ratio.
5-year conditional probability of survival as a function of the
coefficients listed in table 3.
5-year conditional probability of survival as a function of age,
gender, FEV
1
%, weight-for-age z score, pancreatic sufficiency,
diabetes mellitus, Staphylococcus aureus, Burkerholderia cepacia,
and number of acute exacerbations (coefficients not shown).
5-year conditional probability of survival as a function of 22
terms, including interaction terms in addition to the terms listed in
table 3. Stepwise procedures eliminate all but one interaction term
to produce the best model reported in table 3 (and evaluated in the
top row of this table).
5-year conditional probability of survival as a function of age,
gender, and forced expiratory volume in 1 second as a percentage
of predicted normal (FEV
1
%). The model is based on variables
identified by Kerem et al. (15).
# Refer to reference 15.
1,460
1,462
1,472
1,636
1,635
Hosmer-
Lemeshow
p value*
0.54
0.99
0.79
0.007
0.014

2
likelihood
ratio
350 Liou et al.
Am J Epidemiol Vol. 153, No. 4, 2001
FIGURE 1. Graphic representation of application of the Hosmer-Lemeshow test (31) to validate a 5-year logistic regression model of cystic
fibrosis, United States, 1993. The difference in predictive ability between the two models shown was statistically significant (p < 0.0001). A.
Comparison of actual and predicted 5-year survival for patients in the validation set (table 2). The percentage of patients predicted to survive
was calculated directly by using the new model (table 3). A perfect fit of predictions to actual outcomes would produce the dashed line shown,
with slope 1 and intercept 0. The smoothed-curve fit of our model (44, 45) closely approximated this ideal. B. Comparison of actual and pre-
dicted percentage of survival using a 5-year logistic regression model based on age, gender, and forced expiratory volume in 1 second as a
percentage of predicted normal (15). This model did not identify any patients with the lowest chances of survival, so the curve abruptly ends
when the percentage predicted to survive drops below approximately 30%. The model was inaccurate when the percentage predicted to sur-
vive 5 years within the group was less than 60%, as shown by the large deviation of the model curve from ideal (dashed line).
Five-Year Survivorship Model of Cystic Fibrosis 351
Am J Epidemiol Vol. 153, No. 4, 2001
enable the weight-for-age z score and the conditional prob-
ability of 5-year survival to be calculated rapidly. These
worksheets (available on the Journal website at
www.jhsph.edu/Publications/JEPI/liou.htm) reduce the
chance of arithmetic errors during calculation. Because they
eliminate the need to calculate the exponential term inherent
in a logistic regression model, these worksheets can be used
when the only tools available are paper and pencil.
Our survivorship model of cystic fibrosis may have simi-
lar widespread applications. It may help to improve study
design by providing a way to select patients with equivalent
survival predictions for control and experimental arms of
prospective studies. The model also may be useful in evalu-
ating whether survival has changed in response to changing
practice patterns. As a research tool, it has the potential for
use in investigating the effect of therapies that have not been
evaluated randomly, such as lung transplantation. Our
analyses suggest that many patients who undergo transplan-
tation have no survival benefit from the procedure. (T. G.
Liou et al., University of Utah, unpublished manuscript).
Our model will be an objective aid to individual cystic fibro-
sis patients and their physicians contemplating difficult ther-
apeutic choices.
ACKNOWLEDGMENTS
This work was supported in part by a grant from the
Margolis Family Foundation of Utah and by the Cystic
Fibrosis Foundation, Bethesda, Maryland.
The authors are grateful to Ase Sewall, Samia
Buckingham, Bernie LaSalle, Dr. Steve Alder, Monica
Brooks, and Pieter Bowman for technical assistance. They
thank Josh Cherry, Dr. John Hoidal, Dr. John Michael, Dr.
Wayne Samuelson, Dr. Preston Campbell, Dr. Barbara
Chatfield, and Dr. Mike Kosorok for valuable suggestions
and reviews of the manuscript, and they appreciate helpful
suggestions from Dr. George Diaz and Carmen Henshaw
regarding the clinical worksheets. The authors are indebted
to Dr. David Huang for his ready and sound advice on mul-
tiple occasions during the project.
REFERENCES
1. Anderson DH. Cystic fibrosis of the pancreas and its relation
to celiac disease. A clinical and pathologic study. Am J Dis
Child 1938;56:34495.
2. Farber S. Some organic digestive disturbances in early life. J
Michigan Med Sci 1945;44:58794.
3. Cystic Fibrosis Foundation Patient Registry annual data report,
1998. Bethesda, MD: Cystic Fibrosis Foundation, 1999.
4. Isles A, Maclusky I, Corey M, et al. Pseudomonas cepacia
infection in cystic fibrosis: an emerging problem. J Pediatr
1984;104:20610.
5. Thomassen MJ, Demko CA, Klinger JD, et al. Pseudomonas
cepacia colonization among patients with cystic fibrosis. Am
Rev Respir Dis 1985;131:7916.
6. Rodman HM, Doershuk CF, Roland JM. The interaction of 2 dis-
eases: diabetes mellitus and cystic fibrosis. Medicine 1986;65:
38997.
7. Reisman J, Corey M, Canny G, et al. Diabetes mellitus in
patients with cystic fibrosis: effect on survival. Pediatrics
1990;86:3747.
8. Bedrossian CWM, Greenberg SP, Singer DB, et al. The lung in
cystic fibrosis. A quantitative study including prevalence of
pathologic findings among different age groups. Hum Pathol
1976;7:195204.
9. Cohen LF, di SantAgnese PA, Friedlander J. Cystic fibrosis
and pregnancy: a national survey. Lancet 1980;2:8424.
10. Lanng S, Thorsteinsson B, Lund-Andersen C, et al. Diabetes
mellitus in Danish cystic fibrosis patients: prevalence and late
diabetic complications. Acta Paediatr 1994;83:727.
11. FitzSimmons SC, Fitzpatrick S, Thompson B, et al. Alongitu-
dinal study of the effects of pregnancy on 325 women with
cystic fibrosis. Pediatr Pulmonol 1996;suppl 13:99101.
12. Davis PB, Byard PJ, Konstan MW. Identifying treatments that
halt progression of pulmonary disease in cystic fibrosis.
Pediatr Res 1997;41:1615.
13. Lai HC, Kosorok MR, Sondel SA, et al. Growth status in chil-
dren with cystic fibrosis based on the National Cystic Fibrosis
Patient Registry data: evaluation of various criteria used to
identify malnutrition. J Pediatr 1998;132:47885.
14. Rosenfeld M, Davis R, FitzSimmons S, et al. Gender gap in
cystic fibrosis mortality. Am J Epidemiol 1997;145:794803.
15. Kerem E, Reisman J, Corey J, et al. Prediction of mortality in
patients with cystic fibrosis. N Engl J Med 1992;326:118791.
16. Fuchs HJ, Borowitz DS, Christiansen DH, et al. Effect of
aerosolized recombinant human DNase on exacerbations of
respiratory symptoms and on pulmonary function in patients
with cystic fibrosis. The Pulmozyme Study Group. N Engl J
Med 1994;331:63742.
17. Eigen H, Rosenstein BJ, FitzSimmons S, et al. A multicenter
study of alternate-day prednisone therapy in patients with cys-
tic fibrosis. Cystic Fibrosis Foundation Prednisone Trial
Group. J Pediatr 1995;126:51523.
18. Frederiksen B, Lanng S, Koch C, et al. Improved survival in
the Danish center-treated cystic fibrosis patients: results of
aggressive treatment. Pediatr Pulmonol 1996;21:1538.
19. Mortensen J, Falk M, Groth S, et al. The effects of postural
drainage and positive expiratory pressure physiotherapy on tra-
cheobronchial clearance in cystic fibrosis. Chest 1991;100:
13507.
20. Desmond KJ, Schwenk WF, Thomas E, et al. Immediate and
long-term effects of chest physiotherapy in patients with cys-
tic fibrosis. J Pediatr 1983;103:53842.
21. Massard G, Shennib H, Metras D, et al. Double-lung trans-
plantation in mechanically ventilated patients with cystic
fibrosis. Ann Thorac Surg 1993;55:108792.
22. Aris RM, Gilligan PH, Neuringer IP, et al. The effects of pan-
resistant bacteria in cystic fibrosis patients on lung transplant
outcome. Am J Respir Crit Care Med 1997;155:1699704.
23. Corey M, Edwards L, Levison H, et al. Longitudinal analysis
of pulmonary function decline in patients with cystic fibrosis.
J Pediatr 1997;131:80914.
24. Whiteford ML, Wilkinson JD, McColl JH, et al. Outcome of
Burkholderia (Pseudomonas) cepacia colonisation in children
with cystic fibrosis following a hospital outbreak. Thorax
1995;50:11948.
25. Chase HP, Long MA, Lavin MH. Medical progress: cystic
fibrosis and malnutrition. J Pediatr 1979;95:33747.
26. Lanng S, Thorsteinsson B, Nerup J, et al. Diabetes mellitus in
cystic fibrosis: effect of insulin therapy on lung function and
infections. Acta Paediatr Scand 1994;83:84953.
27. FitzSimmons SC. The changing epidemiology of cystic fibro-
sis. J Pediatr 1993;122:19.
28. Becker RA, Chambers JM, Wilks AR. The new S language.
Rev ed. Boca Raton, FL: CRC Press, 1988.
29. Hankinson JL, Odencrantz JR, Fedan KB. Spirometric refer-
ence values from a sample of the general US population. Am J
Respir Crit Care Med 1999;159:17987.
30. Dibley MJ, Goldsby JB, Staehling NW, et al. Development of
352 Liou et al.
Am J Epidemiol Vol. 153, No. 4, 2001
normalized curves for the international growth reference: histor-
ical and technical considerations. Am J Clin Nutr 1987;46:
73648.
31. Hosmer DW, Lemeshow S. Applied logistic regression. New
York, NY: John Wiley & Sons, Inc, 1989:638, 11518, 1405.
32. Venables WN, Ripley BD. Modern applied statistics with S-
Plus. 2nd ed. New York, NY: Springer-Verlag, 1997.
33. Green MS, Symons MJ. Acomparison of the logistic risk func-
tion and the proportional hazards model in prospective epi-
demiologic studies. J Chronic Dis 1983;36:71524.
34. Abbott RD. Logistic regression in survival analysis. Am J
Epidemiol 1985;121:46571.
35. Jankelson D, Robinson M, Parsons S, et al. Cystic fibrosis and
pregnancy. Aust N Z J Obstet Gynaecol 1998;38:1804.
36. Snell GI, de Hoyos A, Krajden M, et al. Pseudomonas cepacia
in lung transplant recipients with cystic fibrosis. Chest 1993;
103:46671.
37. Ward SA, Tomezsko JL, Holschaw DS, et al. Energy expendi-
ture and substrate utilization in adults with cystic fibrosis and
diabetes mellitus. Am J Clin Nutr 1999;69:91319.
38. Finkelstein SM, Wielinski CL, Elliott GR, et al. Diabetes mel-
litus associated with cystic fibrosis. J Pediatr 1988;112:3737.
39. Knaus WA, Draper EA, Wagner DP, et al. APACHE II: A
severity of disease classification system. Crit Care Med 1985;
13:81829.
40. Amato MB, Barbas CS, Medeiros DM, et al. Beneficial effects
of the open lung approach with low distending pressures in
acute respiratory distress syndrome. Aprospective randomized
study on mechanical ventilation. Am J Respir Crit Care Med
1995;152:183546.
41. Noseworthy TW, Konopad E, Shustack A, et al. Cost account-
ing of adult intensive care: methods and human and capital
inputs. Crit Care Med 1996;24:116872.
42. Carson S, Stocking C, Podsadecki T, et al. Effects of organiza-
tional change in the medical intensive care unit of a teaching
hospital. JAMA 1996;276:3228.
43. Pittet D, Thievent B, Wenzel RP, et al. Bedside prediction of
mortality from bacteremic sepsis. Am J Respir Crit Care Med
1996;153:68493.
44. Friedman JH. A variable span smoother. Stanford, CA:
Department of Statistics, Stanford University, 1984. (Technical
report LCS5).
45. Mathsoft. S-Plus 2000 guide to statistics. Vol 1. Seattle, WA:
MathSoft, Data Analysis Products Division, 2000.